Memory Store

Memory Store is the startup product I work on at Julep AI (Autotelic). The domain is cognitive memory architecture for agents and teams: how durable context is encoded, retrieved, assimilated, and turned into living Briefs — not chat logs that evaporate when the session ends.

Agents forget. Memory Store is the substrate for what the system knows across Slack, Linear, GitHub, docs, meetings, and customer conversations. Typed primitives (checkin, recall, record, plus Briefs, episodes, entities, and threads) give humans and agents one shared workspace world model.

Domain: cognitive memory architecture

The product thesis borrows structure from cognitive science without claiming biological fidelity:

• Schemas / frames — organize company knowledge so new information is interpreted through existing structure (encoding), fits without rewrite (assimilation), or forces an update (accommodation).
• Scripts & cues — retrieval paths that make the right episode or decision findable under load.
• Living Briefs — the customer-legible surface synthesized from memory with provenance; not a hand-edited wiki page that drifts from source truth.
• Attribution — decisions and summaries stay grounded in which sources and workflows produced them.

Near-term scope is deliberate: workspace-scoped intelligence (complete for a company), not pretend-general AGI memory. Public framing often uses the agi.new language; the engineering bar is compounding context under real product constraints.

Operating loop

Sources → Memory Store → World model → Living Briefs / PRDs → Plans / Tasks → Agent runs → Outputs → Human review → Memory updates

Closing the loop with memory updates is what makes the system compound rather than a one-shot pipeline. Paired with a durable task graph, memory covers knowledge while the task layer covers what the system is doing about it.

What exists publicly

• Product: memory.store
• Installable plugins: julep-ai/mem-plugins
• Writing on this site: Memory Store, the intern problem, Memory enabled Claude Code
• Portfolio spine: Cognitive Systems Portfolio Thesis

Boundary with simulation and agent OS work

Memory Store is not the Agora tick brain and not Sansara’s vault runtime. In the agent-simulation architecture, durable memory must not hold per-tick activations or decide reward — that separation is load-bearing. Cross-links:

• Agent Simulation Lab — how Agora measures populations under rules
• Sansara — vault-native agent OS experiments
• RSI Lab — verifier-gated research discipline that provenance-shaped memory enables

Current product direction

The open gap after durable memory and task substrates is supervision UX: spawning, monitoring, and retiring grounded agents without losing the world model. Research claims on this personal site stay scoped to published writing and sealed lab reports — customer internals stay private.